1. INTRODUCTION TO DIGITAL SIGNAL PROCESSORS (DSPs) Prof. Brian L. Evans Contributions by Niranjan Damera-Venkata and Magesh Valliappan Embedded Signal Processing Laboratory The University of Texas at Austin Austin, TX 78712-1084 http://signal.ece.utexas.edu/ Accumulator architecture Load-store architecture Memory-register architecture

2. 2 -2 Outline Signal processing applications Conventional DSP architecture Pipelining in DSP processors RISC vs. DSP processor architectures TI TMS320C6000 DSP architecture introduction Signal processing on general-purpose processors Conclusion

3. 2 -3 Signal Processing Applications Embedded system demand: volume, volume, …  400 Million units/year: automobiles, PCs, cell phones  30 Million units/year: ADSL modems and printers Embedded system cost and input/output rates  Low-cost, medium-throughput: low-end printers, handsets, sound cards, car audio, disk drives  High-cost, high-throughput: high-end printers, wireless basestations, 3-D sonar, 3-D images from 2-D X-rays (tomographic reconstruction) Embedded processor requirements  Inexpensive with small area and volume  Predictable input/output (I/O) rates to/from processor  Power constraints (severe for handheld devices) Single DSP Multiple DSPs

4. 2 -4 Conventional DSP Processors Low cost: $3/processor in volume Deterministic interrupt service routine latency guarantees predictable input/output rates  On-chip direct memory access (DMA) controllers Processes streaming input/output separately from CPU Sends interrupt to CPU when block has been read/written  Ping-pong buffering CPU reads/writes buffer 1 as DMA reads/writes buffer 2 After DMA finishes buffer 2, roles of buffers 1 & 2 switch Low power consumption: 10-100 mW  TI TMS320C54 0.32 mA/MIP  76.8 mW at 1.5 V, 160 MHz  TI TMS320C55 0.05 mA/MIP  22.5 mW at 1.5 V, 300 MHz Based on conventional (pre-1996) architecture

5. 2 -5 Conventional DSP Architecture Multiply-accumulate (MAC) in 1 instruction cycle Harvard architecture for fast on-chip I/O  Data memory/bus separate from program memory/bus  One read from program memory per instruction cycle  Two reads/writes from/to data memory per inst. cycle Instructions to keep pipeline (3-6 stages) full  Zero-overhead looping (one pipeline flush to set up)  Delayed branches Special addressing modes supported in hardware  Bit-reversed addressing (e.g. fast Fourier transforms)  Modulo addressing for circular buffers (e.g. FIR filters)

6. 2 -6 Conventional DSP Architecture (con’t) x N-K+1 x N-K+2 x N-1 xNxN Data Shifting Using a Linear Buffer TimeBuffer contentsNext sample x N+1 x N+3 x N+2 n=N n=N+1 n=N+2 x N-K+3 x N-K+4 x N+1 x N+2 x N-K+2 x N-K+3 xNxN x N+1 Modulo Addressing Using a Circular Buffer Time Buffer contents Next sample n=N n=N+1 n=N+2 x N-2 x N-1 x N-K+1 x N-K+2 x N-K+4 x N+1 x N+2 x N+3 x N-2 x N-1 x N+1 x N-K+2 xNxN x N-2 x N-1 x N+1 x N+2 xNxN xNxN xNxN xNxN x N-K+3 x N-K+4 Buffer of length K  Used in finite and infinite impulse response filters Linear buffer  Sort by time index  Update: discard oldest data, copy old data left, insert new data Circular buffer  Oldest data index  Update: insert new data at oldest index, update oldest index

7. 2 -7 Conventional DSP Processors Summary

8. 2 -8 Conventional DSP Processor Families Floating-point DSPs  Used in initial prototyping of algorithms  Resurgence due to professional and car audio Different on-chip configurations in each family  Size and map of data and program memory  A/D, input/output buffers, interfaces, timers, and D/A Drawbacks to conventional DSP processors  No byte addressing (needed for images and video)  Limited on-chip memory  Limited addressable memory on fixed-point DSPs (exceptions include Motorola 56300 and TI C5409)  Non-standard C extensions for fixed-point data type DSP Market Fixed-point 95% Floating-point 5%

9. 2 -9 Pipelining Process instruction stream in stages (like oil flows in an oil pipeline) Increase throughput Managing Pipelines Compiler or programmer Pipeline interlocking Sequential (Motorola 56000) Pipelined (Most conventional DSPs) Superscalar (Pentium, MIPS) Superpipelined (TMS320C6000) Fetch Read ExecuteDecode FetchDecodeRead Execute FetchRead ExecuteDecode FetchRead ExecuteDecode

10. 2 -10 Time-stationary pipeline model  Programmer controls each cycle  Example: Motorola DSP56001 (has separate X/Y data memories/registers) Data-stationary pipeline model  Programmer specifies data operations  Example: TI TMS320C30 Interlocked pipeline  “Protection” from pipeline effects  May not be reported by simulators: inner loops may take extra cycles Pipelining: Operation MAC X0,Y0,A X:(R0)+,X0 Y:(R4)-,Y0 MPYF *++AR0(1),*++AR1(IR0),R0 DEFGHIJKLLDEFGHIJKLL CDEFGHIJK-LCDEFGHIJK-L BCDEFGHIJK-LBCDEFGHIJK-L ABCDEFGHIJK-LABCDEFGHIJK-L FDRE Execute ReadDecodeFetch MAC means multiplication-accumulation.

11. 2 -11 A control hazard occurs when a branch instruction is decoded  Processor “flushes” the pipeline, or  Use delayed branch (expose pipeline) A data hazard occurs because an operand cannot be read yet  Intended by programmer, or  Interlock hardware inserts “bubble”  TI TMS320C5000 (20 CPU & 16 I/O registers, one accumulator, and one address pointer ARP implied by * ) Pipelining: Hazards LAR AR2, ADDR ; load address reg. LACC *- ; load accumulator w/ ; contents of AR2 D E F br G - X Y Z FDRE Execute ReadDecodeFetch C D E F br - X - Y Z B C D E F br - X - Y Z A B C D E F br - X - Y Z LAR: 2 cycles to update AR2 & ARP; need NOP after it

12. 2 -12 A repeat instruction repeats one instruction or a block of instructions after repeat The pipeline is filled with repeated instruction (or block of instructions) Cost: one pipeline flush only Pipelining: Avoiding Control Hazards ; repeat TBLR inst. COUNT-1 times RPT COUNT TBLR *+ High throughput performance of DSPs is helped by on-chip dedicated logic for looping (downcounters/looping registers) D E F rpt X FDRE Execute ReadDecodeFetch C D E F rpt - X B C D E F rpt - X AB C D E F rpt - X

13. 2 -13 RISC vs. DSP: Instruction Encoding RISC: Superscalar, out-of-order execution DSP: Horizontal microcode, in-order execution Reorder Load/store Integer Unit Floating-Point Unit Load/store Address MultiplierALU Memory

14. 2 -14 RISC vs. DSP: Memory Hierarchy RISC DSP Registers Out of order I/D Cache Physical memory TLB Registers DMA Controller I Cache Internal memories External memories TLB: Translation Lookaside Buffer DMA: Direct Memory Access

15. 2 -15 Program RAM Data RAM or Cache Internal Buses Control Regs Regs (B0-B15)Regs (A0-A15).D1.M1.L1.S1.D2.M2.L2.S2 CPU Addr Data External Memory -Sync -Async DMA Serial Port Host Port Boot Load Timers Pwr Down TI TMS320C6000 DSP Architecture Simplified Architecture

16. 2 -16 TI TMS320C6000 DSP Architecture Families: All support same C6000 instruction set C6200 fixed-pt. 150- 300 MHz ADSL, printers C6400 fixed pt. 300-1,000 MHz video, wireless basestations C6700 floating 100- 225 MHz medical imaging, pro-audio TMS320C6211: 150 MHz, $21 in volume 300 million multiply-accumulates/s, 1200 RISC MIPS On-chip memory: 16 kwords program, 16 kwords data TMS320C6701 Evaluation Module Board 167 MHz 334 million multiply-accumulates/s, 1336 RISC MIPS On-chip memory: 16 kwords program, 16 kwords data External: one 133-MHz 64-kword, two 100-MHz 1-Mword

17. 2 -17 TI TMS320C6000 DSP Architecture Very long instruction word (VLIW) size of 256 bits  Eight 32-bit functional units with single cycle throughput  One instruction cycle per clock cycle Data word size is 32 bits  16 (32 on C64) 32-bit registers in each of two data paths  40 bits can be stored in adjacent even/odd registers Two parallel data paths  Data unit - 32-bit address calculations (modulo, linear)  Multiplier unit - 16 bit  16 bit with 32-bit result  Logical unit - 40-bit (saturation) arithmetic & compares  Shifter unit - 32-bit integer ALU and 40-bit shifter

18. 2 -18 TI TMS320C6000 Instruction Set.S Unit ADDNEG ADDKNOT ADD2OR ANDSET BSHL CLRSHR EXTSSHL MVSUB MVCSUB2 MVKXOR MVKHZERO.L Unit ABS NOT ADD OR AND SADD CMPEQ SAT CMPGT SSUB CMPLT SUB LMBD SUBC MV XOR NEG ZERO NORM.M Unit MPYSMPY MPYHSMPYH.D Unit ADD ST ADDA SUB LD SUBA MV ZERO NEG Other NOPIDLE C6000 Instruction Set by Functional Unit Six of the eight functional units can perform integer add, subtract, and move operations

19. 2 -19 TI TMS320C6000 Instruction Set Arithmetic ABS ADD ADDA ADDK ADD2 MPY MPYH NEG SMPY SMPYH SADD SAT SSUB SUB SUBA SUBC SUB2 ZERO Logical AND CMPEQ CMPGT CMPLT NOT OR SHL SHR SSHL XOR Bit Management CLR EXT LMBD NORM SET Data Management LD MV MVC MVK MVKH ST Program Control B IDLE NOP C6000 Instruction Set by Category (un)signed multiplication saturation/packed arithmetic

20. 2 -20 C6000 vs. C5000 Addressing Modes ADD #0FFh add.L1 -13,A1,A6 TI C5000TI C6000 (implied) add.L1 A7,A6,A7 ADD 010h not supported ADD * ldw.L1 *A5++[8],A1 Immediate  The operand is part of the instruction Register  Operand is specified in a register Direct  Address of operand is part of the instruction (added to imply memory page) Indirect  Address of operand is stored in a register

21. 2 -21 TI TMS320C6000 DSP Architecture Deep pipeline  7-11 stages in C6200: fetch 4, decode 2, execute 1-5  7-16 stages in C6700: fetch 4, decode 2, execute 1-10  Pentium IV has an estimated 20 pipeline stages Avoid using branch instructions in code  Branch instruction in pipeline disables interrupts: latency of a branch is 5 cycles  Avoid branches by using conditional execution: every instruction can be conditionally executed No hardware protection against pipeline hazards  Compiler and assembler must prevent pipeline hazards

22. 2 -22 TI TMS320C6700 Extensions.S Unit ABSDP CMPLTSP ABSSP RCPDP CMPEQDP RCPSP CMPEQSP RSARDP CMPGTDP RSQRSP CMPGTSP SPDP CMPLTDP.L Unit ADDDP INTSP ADDSP SPINT DPINT SPTRUNC DPSP SUBDP DPTRUNC SUBSP INTDP.M Unit MPYDP MPYID MPYI MPYSP.D Unit ADDAD LDDW C6700 Floating Point Extensions by Unit Four functional units can perform IEEE single-precision (SP) and double-precision (DP) floating-point add, subtract, move. Operations beginning with R are reciprocal calculations.

23. 2 -23 C6712 vs. C6713 C6712 150 MHz clock, 900 MFLOPS 4 kB/4kB of L1 program/data memory 64 kB of L2 cache 0 kB of L2 SRAM 1200 MB/s on-chip data bus bandwidth $13.50 each in volume C6713 225 MHz clock, 1350 MFLOPS 4 kB/4kB of L1 program/data memory 256 kB of L2 cache 192 kB of L2 SRAM 1800 MB/s on-chip data bus bandwidth $26.85 each in volume Information as of December 14, 2003

24. 2 -24 Digital Signal Processor Cores ASIC with  Programmable digital signal processor core  RAM  ROM  Standard cells  Codec  Peripherals  Gate array  Microcontroller core

25. 2 -25 General Purpose Processors Multimedia applications on PCs  Video, audio, graphics and animation  Repetitive parallel sequences of instructions Single Instruction Multiple Data (SIMD)  One instruction acts on multiple data in parallel  Well-suited for graphics Native signal processing extensions use SIMD  Sun Visual Instruction Set [1995] (UltraSPARC 1/2)  Intel MMX [1996] (Pentium I/II/III/IV)  Intel Streaming SIMD Extensions (Pentium III)

26. 2 -26 DSP on General Purpose Processors (con’t) Programming is considerably tougher  Ability of compilers to generate code for instruction set extensions may lag (e.g. four years for Pentium MMX)  Libraries of routines using native signal processing  Hand code in assembly for best performance Single-instruction multiple-data (SIMD) approach  Pack/unpack data not aligned on SIMD word boundaries  Saturation arithmetic in MMX; not supported in VIS  Extended-precision accumulation in MMX; none in VIS Application speedup for Intel MMX and Sun VIS  Signal and image processing: 1.5:1 to 2:1  Graphics: 4:1 to 6:1 (no packing/unpacking)

27. 2 -27 Intel MMX Instruction Set 64-bit SIMD register (4 data types)  64-bit quad word  Packed byte (8 bytes packed into 64 bits)  Packed word (4 16-bit words packed into 64 bits)  Packed double word (2 double words packed into 64 bits) 57 new instructions  Pack and unpack  Add, subtract, multiply, and multiply/accumulate Saturation and wraparound arithmetic Maximum parallelism possible  8:1 for 8-bit additions  4:1 for 8  16 multiplication or 16-bit additions

28. 2 -28 Concluding Remarks Conventional digital signal processors  High performance vs. power consumption/cost/volume  Excel at one-dimensional processing  Per cycle: 1 16  16 MAC & 4 16-bit RISC instructions TMS320C6000 VLIW DSP family  High performance vs. cost/volume  Excel at multidimensional signal processing  Per cycle: 2 16  16 MACs & 4 32-bit RISC instructions Native signal processing  Available on desktop computers  Excels at graphics  Per cycle: 2 8  16 MACs OR 8 8-bit RISC instructions Use assembly for computational kernels and C for main program (control code, interrupt def.)

29. 2 -29 Concluding Remarks Digital signal processor market 40% annual growth 1990-2000: #1 in semiconductor market Revenue: $4.4B ‘99, $6.1B ‘00, $4.5B ‘01, $4.9B ‘02, $6B ‘03 2000: 44% TI, 23% Agere, 13% Motorola, 10% Analog Dev. 2001: 40% TI, 16% Agere, 12% Motorola, 8% Analog Dev. 2002: 43% TI, 14% Motorola, 14% Agere, 9% Analog Dev. Independent processor benchmarking by industry  Berkeley Design Technology Inc. http://www.bdti.com  EDN Embedded Microprocessor Benchmark Consortium http://www.eembc.org Web resources  Newsgroup comp.dsp: FAQ http://www.bdti.com/faq/dsp_faq.html  Embedded processors and systems: http://www.eg3.com  On-line courses: http://www.techonline.com

30. 2 -30 References  G. E. Allen, B. L. Evans, and D. C. Schanbacher, “Real-Time Sonar Beamforming on a Unix Workstation,” Proc. IEEE Asilomar Conf. On Signals, Systems, and Computers, pp. 764-768, 1998. http://www.ece.utexas.edu/~bevans/papers/1998/beamforming/  R. Bhargava, R. Radhakrishnan, B. L. Evans, and L. K. John, “Evaluating MMX Technology Using DSP and Multimedia Applications,” Proc. IEEE Sym. On Microarchitecture, pp. 37-46, 1998. http://www.ece.utexas.edu/~ravib/mmxdsp/  W. Chen, H. J. Reekie, S. Bhave, and E. A. Lee, “Native Signal Processing on the UltraSPARC in the Ptolemy Environment,” Proc. IEEE Asilomar Conf. On Signals, Systems, and Computers, 1996. http://www.ece.utexas.edu/~bevans/courses/ee382c/lectures/21_nsp/vis/  B. L. Evans, “EE345S Real-Time DSP Laboratory,” UT Austin. http://www.ece.utexas.edu/~bevans/courses/realtime/  B. L. Evans, “EE382C-9 Embedded Software Systems,” UT Austin. http://www.ece.utexas.edu/~bevans/courses/ee382c/  A. Kulkarni and A. Dube, “Evaluation of the Code Generation Domain in Ptolemy,” http://www.ece.utexas.edu/~bevans/talks/benchmarking97/sld001.htm  P. Lapsley, J. Bier, A. Shoham, and E. A. Lee, DSP Processor Fundamentals, IEEE Press, 1997.